New Query Engine for AI Agent Data Analysis
A new paper on arXiv (2605.27785) proposes a query engine designed for analyzing unstructured text from AI agents, such as traces, chat logs, and reasoning chains. The authors argue that traditional SQL cannot answer questions like "show me where the agent got confused" because text requires a model in the query path. The engine targets client-side AI applications like Claude Code, Cursor, and Claude Desktop, which run in JavaScript runtimes and need a lightweight, JS-native distribution that can be bundled inside a cold start. The engine must be small enough to ship within such applications, addressing the difficulty of using lakehouse read paths (Spark, Trino, managed warehouses) from JS runtimes.
Key facts
- arXiv paper 2605.27785 proposes a query engine for unstructured text from AI agents.
- Traditional SQL cannot answer queries about agent confusion because text needs a model in the query path.
- Target applications include Claude Code, Cursor, Claude Desktop, and in-browser agents.
- The engine must be JS-native and small enough to ship inside a cold start.
- Lakehouse read paths like Spark, Trino, and managed warehouses are not suitable for JS runtimes.
- The fastest-growing data in production is unstructured text from agents.
- People want to analyze agent traces, chat logs, reasoning chains, and model outputs.
- The engine is designed for client-side AI applications that host both human and LLM agent in the same process.
Entities
Institutions
- arXiv